A street view space multi-modal fusion intersection scene safety risk quantification method

Abstract

The application discloses a kind of street view space multimodal fusion intersection scene safety risk quantification method, based on historical traffic accident data and considering distance attenuation effect, the traffic accident occurrence intensity index of intersection scene area is calculated and as training label data;Based on street view panoramic image and road network data, the visual features of intersection scene are calculated, the spatial structure features of intersection scene are calculated;Based on interest point and traffic flow data, the semantic features of intersection scene are calculated;The above-mentioned features are combined to construct intersection scene safety risk feature vector;Intensity index and safety risk feature vector are combined, and the safety risk quantification model of intersection scene is constructed using random forest regression algorithm.The application designs traffic accident occurrence intensity index, and models are considered from the visual, spatial structure and semantic dimension features of intersection scene, which has low dependence on computing power and label data, strong operability, and can provide technical support for road safety risk early warning and safe city construction.

Description

Technical Field

[0001] This invention relates to the field of road safety technology, specifically to a method for quantifying safety risks in intersection scenarios based on multimodal fusion of street view space. Background Technology

[0002] The quantitative calculation of road safety risk indicators requires consideration of numerous factors, including the urban physical environment, built environment, and traffic conditions (including people, vehicles, and logistics), making it a challenging issue in the fields of urban perception and intelligent transportation. Traditional methods rely on expert scoring and historical traffic accident data to assess safety risks in localized high-accident areas. However, existing methods based on machine learning and manual feature extraction lack consideration for the multimodal semantic information of urban space and depend on labeled data and indicator feature data, making them unsuitable for situations where labeled data and indicator feature data are limited or missing.

[0003] Intersections are a typical feature of urban roads, a crucial location for navigation and other tasks, and also a high-risk area. It is known that built environment characteristics, road surface characteristics, road network structure characteristics, and traffic flow characteristics are all major factors influencing intersection safety risks. Therefore, multi-source urban observation data and geographic data can be used to characterize intersection scenarios and support the construction of quantitative models of intersection safety risks. Summary of the Invention

[0004] This invention relates to a method for quantifying safety risks in intersection scenarios based on multimodal fusion of street view space. It can quantify and calculate safety risk indicators for intersection scenarios, providing basic data support for road safety risk assessment, traffic safety risk early warning, and low-risk navigation path planning.

[0005] The present invention specifically includes the following steps:

[0006] Step 1: Based on historical traffic accident data and considering the distance decay effect, calculate the traffic accident intensity index for the intersection scene area;

[0007] Step 2: Extracting multimodal information from the intersection scene area. The specific steps are as follows:

[0008] Step 1. Based on street view panoramic images and road network data, calculate the visual features of the intersection scene area, including visual openness and road semantic area under panoramic projection; the visual openness of the intersection scene is used to characterize the visual visibility or visibility of the intersection area, and the road semantic area index under panoramic projection is used to characterize the road width of the intersection area.

[0009] Step 2. Based on road network data, calculate the spatial structure characteristics of the intersection scene area, including the number of road branches at the intersection and the road network integration index;

[0010] Step 3. Based on points of interest and traffic flow data, calculate the semantic features of the intersection scene area, including the density of points of interest and the intersection traffic flow index;

[0011] Step 3: Use function normalization to process the visual features, spatial structure features, and semantic features obtained in Step 2 to construct a safety risk feature vector for the intersection scene;

[0012] Step 4: Based on the traffic accident intensity index and the intersection scenario safety risk feature matrix, a quantitative model of intersection safety risk is established using the random forest regression algorithm to obtain the safety risk index of the target intersection scenario.

[0013] Step one specifically involves:

[0014] First, calculate the historical traffic accident TA. ij The location of the incident is near the nearest intersection RI i Road network distance Dis TAij,RIi ;

[0015] Then, the historical traffic accident TA is calculated based on the Gaussian kernel function. ij For the nearest intersection RI i Influence factors :

[0016] ;

[0017] In the formula: ∈{ , …, } represents the nearest intersection RI i It contains data on j historical traffic accidents; This indicates the preset intersection buffer zone radius.

[0018] Finally, calculate the nearest intersection RI. i Traffic accident intensity TASi: .

[0019] Step 1 specifically involves:

[0020] First, a semantic segmentation map of the street view panoramic image is obtained. Then, an equal-area cylindrical projection model is used to project the semantic segmentation map, and the set of sky semantic pixels in the reprojected semantic segmentation map is calculated. Pano, a calculation of street view panoramic images ik Location and nearest intersection RI i Distance attenuation factor :

[0021] ;

[0022] In the formula: ∈{ , …, } represents the nearest intersection RI i It contains k panoramic street view images. Indicates the preset intersection buffer zone radius;

[0023] The semantic segmentation map is a street view panoramic image that has undergone semantic segmentation processing; the reprojection semantic segmentation map is a semantic segmentation map processed by an equal-area cylindrical projection model.

[0024] Then, calculate the visual openness. :

[0025] ;

[0026] In the formula: This represents the set of sky semantic pixels in the reprojected semantic segmentation map. Indicates the resolution of the reprojected semantic segmentation map;

[0027] Finally, the road semantic pixel set in the reprojected semantic segmentation map is calculated. Calculate the semantic area of ​​roads under panoramic projection. :

[0028] ;

[0029] In the formula: This represents the set of road semantic pixels in the reprojected semantic segmentation map. This represents the resolution of the reprojected semantic segmentation map.

[0030] Step 2 specifically involves:

[0031] Calculate the number of road branches at each intersection based on road network data. Number of road branches at the intersection For: the nearest intersection RI i Number of intersecting road segments; Calculation of nearest neighbor intersection RI based on space syntax i Integration index ;

[0032] Step 3 specifically involves:

[0033] First, based on the traffic flow volume, semantic indicators of traffic flow per unit time are segmented for each intersection;

[0034] Then, the traffic flow per target unit of time is selected as the traffic flow index. The specific calculation method is: the amount of traffic passing through the intersection per target unit of time;

[0035] Then, interest point data with insignificant semantic information is filtered out. The specific process is as follows: Based on the industry classification directory and classification attributes of interest points, interest point data of the following categories are filtered out, including: addresses, natural features, and administrative landmarks;

[0036] Finally, calculate the Points of Interest (POIs). im Location and nearest intersection RI i Distance attenuation factor :

[0037] ;

[0038] In the formula: ∈{ , …, } represents the nearest intersection RI i There are m points of interest. This indicates the preset intersection buffer zone radius.

[0039] Calculate the density of interest points : .

[0040] Step three specifically involves:

[0041] First, all visual features, spatial structure features, and semantic features are normalized using the min-max function;

[0042] Then, construct the intersection scenario safety risk feature vector X. RIi ;

[0043] ;

[0044] In the formula, the symbol This represents the characteristics after normalization. For visual openness, For the semantic area of ​​the road, As an integration index, This represents the density of interest points.

[0045] Step four specifically involves:

[0046] First, the dataset is combined and divided into training and testing sets. The dataset consists of traffic accident intensity indicators and intersection scene safety risk matrices. Then, the random forest regression algorithm is used for training to obtain a quantitative model of intersection scene safety risks.

[0047] The beneficial effects of this invention are:

[0048] First, this invention comprehensively considers visual information, spatial structure information, and semantic information that affect the safety risks of intersections in street scene scenarios, and can describe safety risk factors more comprehensively and accurately.

[0049] Secondly, this invention utilizes historical traffic accident data to indirectly describe safety risks, thus addressing to some extent the problem of missing label data in road safety risk modeling tasks. Furthermore, it designs a traffic accident intensity index that considers distance attenuation effects, indirectly and effectively quantifying safety risk factors in intersection areas.

[0050] Third, this invention utilizes low-computing-power, lightweight machine learning methods for modeling and makes full use of various types of open-source urban observation data and open-source geographic data, thus possessing strong portability and operability. Attached Figure Description

[0051] Figure 1 A flowchart illustrating a method for quantifying safety risks in intersection scenarios through multimodal fusion of street view space;

[0052] Figure 2 This is a schematic diagram of multimodal feature extraction in the intersection scene area of ​​this invention. Detailed Implementation

[0053] The present invention will now be further described with reference to the accompanying drawings.

[0054] like Figure 1 As shown, a method for quantifying the safety risks of intersection scenes based on multimodal fusion of street scene space is proposed. The method uses the historical traffic accident intensity (TASi) index to indirectly characterize the safety risks of the intersection area and uses TASi as label data for modeling. The nodes of the road network are regarded as the intersection center to identify the intersection. The intersection area is defined as a buffer zone with the intersection as the center and buf as the radius.

[0055] Specifically, the following steps are included:

[0056] Step 1: Based on historical traffic accident data and considering the distance attenuation effect, calculate the traffic accident intensity index for the intersection scene area; specifically:

[0057] First, calculate the historical traffic accident TA. ij The location of the incident is near the nearest intersection RI i Road network distance Dis TAij,RIi ;

[0058] Then, the historical traffic accident TA is calculated based on the Gaussian kernel function. ij For the nearest intersection RI i Influence factors , :

[0059] ;

[0060] In the formula: ∈{ , …, } represents the nearest intersection RI i It contains j historical traffic accident data; buf is the preset intersection buffer radius;

[0061] Finally, calculate the nearest intersection RI. i Traffic accident intensity TASi: .

[0062] Step 2: Extracting multimodal information from the road scene area, such as... Figure 2 As shown, specifically:

[0063] Step 1. Based on street view panoramic imagery and road network data, calculate the visual features of the intersection scene area, including visual openness and road semantic area under panoramic projection; specifically:

[0064] Obtain the semantic segmentation map from the street view panoramic image, then perform a projection transformation on the semantic segmentation map using an equal-area cylindrical projection model, and calculate the sky semantic pixel set in the reprojected semantic segmentation map. Pano, a calculation of street view panoramic images ik Location and nearest intersection RI i Distance attenuation factor :

[0065] ;

[0066] In the formula: ∈{ , …, } represents the nearest intersection RI i It contains k panoramic street view images. Indicates the preset intersection buffer zone radius;

[0067] Then, calculate the visual openness. :

[0068] ;

[0069] In the formula: This represents the set of sky semantic pixels in the reprojected semantic segmentation map. Indicates the resolution of the reprojected semantic segmentation map;

[0070] Calculate the road semantic pixel set in the reprojection semantic segmentation map Calculate the semantic area of ​​roads under panoramic projection. :

[0071] ;

[0072] In the formula: This represents the set of road semantic pixels in the reprojected semantic segmentation map. This represents the resolution of the reprojected semantic segmentation map.

[0073] Step 2. Based on road network data, calculate the spatial structure characteristics of the intersection scene area, including the number of road branches at the intersection and the road network integration index; specifically:

[0074] Calculate the number of road branches at each intersection based on road network data. Number of road branches at the intersection For: the nearest intersection RI i The number of intersecting road segments;

[0075] RI calculation of nearest neighbor intersections based on space syntax i Integration index Integration index The calculation formula is:

[0076] ;

[0077] In the formula: m represents the total number of road network nodes. Indicates the average depth value;

[0078] Step 3. Based on points of interest and traffic flow data, calculate the semantic features of the intersection scene area, including point of interest density and intersection traffic flow index; specifically:

[0079] First, based on the traffic flow volume, semantic indicators of traffic flow per unit time are segmented for each intersection;

[0080] Then, the traffic flow per target unit of time is selected as the traffic flow index. The specific calculation method is: the amount of traffic passing through the intersection per target unit of time;

[0081] Then, interest point data with insignificant semantic information is filtered out. The specific process is as follows: Based on the industry classification directory and classification attributes of interest points, interest point data in the following categories are filtered out, including: addresses, natural features, and administrative landmarks, as shown in Table 1:

[0082] Table 1 Directory of Common Points of Interest with Insignificant Semantic Information

[0083] Primary Classification subclass Address Address, Residential / Building Number, Commercial Address Artificial features Bridges, overpasses, roundabouts, tunnels Natural features Mountains, islands, lakes, and waterways Administrative Landmark Province / autonomous region, prefecture-level city, district / county, township, village, business district

[0084] Finally, calculate the Points of Interest (POIs). im Location and nearest intersection RIi Distance attenuation factor :

[0085] ;

[0086] In the formula: ∈{ , …, } represents the nearest intersection RI i There are m points of interest. This indicates the preset intersection buffer zone radius.

[0087] Calculate the density of interest points : .

[0088] Step 3: Normalize the visual features, spatial structure features, and semantic features obtained in Step 2 using the min-max function to construct a safety risk feature vector for the intersection scene; specifically:

[0089] First, all visual features, spatial structure features, and semantic features are normalized using the min-max function. The formula for calculating the min-max function is as follows:

[0090] ;

[0091] In the formula: Indicates the nearest intersection RI i One of the characteristics, This represents the minimum value of this characteristic. This represents the maximum value of this characteristic;

[0092] Then, construct the intersection scenario safety risk feature vector X. RIi X RIi The calculation formula is:

[0093] ;

[0094] In the formula, the symbol This represents the characteristics after normalization.

[0095] Step 4: Based on the traffic accident intensity index and the intersection scene safety risk feature matrix, a quantitative model of intersection safety risk is established using the random forest regression algorithm to obtain the safety risk index of the target intersection scene. Specifically: The historical traffic accident intensity index obtained in Step 1 and the intersection scene safety risk feature vector obtained in Step 3 are combined into a dataset, which is then divided into two parts: 80% as the training set and 20% as the test set. The random forest regression algorithm is used for training, with max_features set to 6, to obtain the quantitative model of intersection scene safety risk.

Claims

1. A method for quantifying safety risks in intersection scenes using multimodal fusion of street view space, characterized in that, Specifically, the steps include the following: Step 1: Based on historical traffic accident data and considering the distance decay effect, calculate the traffic accident intensity index of the intersection scene area; use the historical traffic accident intensity TASi index to indirectly characterize the safety risk of the intersection area, and use TASi as label data for modeling; regard the nodes of the road network as the intersection center to identify the intersection, and the intersection area is: a buffer zone with the intersection as the center and buf as the radius. Step 2: Extracting multimodal information from the intersection scene area. The specific steps are as follows: Step 1. Based on street view panoramic images and road network data, calculate the visual features of the intersection scene area, including visual openness and road semantic area under panoramic projection; the visual openness of the intersection scene is used to characterize the visual visibility or visibility of the intersection area, and the road semantic area index under panoramic projection is used to characterize the road width of the intersection area. Step 2. Based on road network data, calculate the spatial structure characteristics of the intersection scene area, including the number of road branches at the intersection and the road network integration index; specifically: calculate the number of road branches at each intersection based on road network data. Number of road branches at the intersection For: the nearest intersection RI i The number of intersecting road segments; RI calculation of nearest intersections based on space syntax i Integration index ; Step 3. Based on points of interest and traffic flow data, calculate the semantic features of the intersection scene area, including the density of points of interest and the intersection traffic flow index; Step 3: Normalize the visual features, spatial structure features, and semantic features obtained in Step 2 using the min-max function, and construct a safety risk feature vector for the intersection scene; Step 4: Based on the traffic accident intensity index and the intersection scenario safety risk feature matrix, a quantitative model of intersection safety risk is established using the random forest regression algorithm to obtain the safety risk index of the target intersection scenario.

2. The method for quantifying intersection scene safety risks through multimodal fusion of street view space according to claim 1, characterized in that, Step one specifically involves: First, calculate the historical traffic accident TA. ij The location of the incident is near the nearest intersection RI i Road network distance Dis TAij,RIi ; Then, the historical traffic accident TA is calculated based on the Gaussian kernel function. ij For the nearest intersection RI i Influence factors : ； In the formula: ∈{ , …, } represents the nearest intersection RI i It contains data on j historical traffic accidents; Finally, calculate the nearest intersection RI. i Traffic accident intensity TASi: .

3. The method for quantifying intersection scene safety risks through multimodal fusion of street view space according to claim 2, characterized in that: Step 1 specifically involves: First, a semantic segmentation map of the street view panoramic image is obtained. Then, an equal-area cylindrical projection model is used to project the semantic segmentation map, and the set of sky semantic pixels in the reprojected semantic segmentation map is calculated. Pano, a calculation of street view panoramic images ik Location and nearest intersection RI i Distance attenuation factor : ； In the formula: ∈{ , …, } represents the nearest intersection RI i It contains k panoramic street view images. Indicates the preset intersection buffer zone radius; Then, calculate the visual openness. : ； In the formula: This represents the set of sky semantic pixels in the reprojected semantic segmentation map. Indicates the resolution of the reprojected semantic segmentation map; Finally, the road semantic pixel set in the reprojected semantic segmentation map is calculated. Calculate the semantic area of ​​roads under panoramic projection. : ； In the formula: This represents the set of road semantic pixels in the reprojected semantic segmentation map. This represents the resolution of the reprojected semantic segmentation map.

4. The method for quantifying intersection scene safety risks through multimodal fusion of street view space according to claim 3, characterized in that: Step 3 specifically involves: First, based on the traffic flow volume, calculate the semantic indicators of traffic flow at each intersection; Then, the traffic flow per target unit of time is selected as the traffic flow index. The specific calculation method is: the amount of traffic passing through the intersection per target unit of time; Then, interest point data with insignificant semantic information is filtered out. The specific process is as follows: Based on the industry classification directory and classification attributes of interest points, interest point data of the following categories are filtered out, including: addresses, natural features, and administrative landmarks; Finally, calculate the Points of Interest (POIs). im Location and nearest intersection RI i Distance attenuation factor : ； In the formula: ∈{ , …, } represents the nearest intersection RI i There are m points of interest. Indicates the preset intersection buffer zone radius; Calculate the density of interest points , The calculation formula is: .

5. The method for quantifying intersection scene safety risks through multimodal fusion of street view space according to claim 4, characterized in that: Step three specifically involves: First, all visual features, spatial structure features, and semantic features are normalized using the min-max function; Then, construct the intersection scenario safety risk feature vector X. RIi X RIi The calculation formula is: ； In the formula, the symbol This represents the characteristics after normalization.

6. The method for quantifying intersection scene safety risks through multimodal fusion of street view space according to claim 1, characterized in that: Step four specifically involves: First, the dataset is combined and divided into training and testing sets. The dataset consists of traffic accident intensity indicators and intersection scene safety risk matrices. Then, the random forest regression algorithm is used for training to obtain a quantitative model of intersection scene safety risks.

Images